Replicating earlier work on mvpa. Try not to overly complicate it--the main point is just to verify we get similar results on a different package to validate prior work. ANd we are primarily interested in validating the very high cross-validation results I got with nltools. Should aim for readable code.

Version 5 uses scikit-learn directly, bypassing mvpa2's framework altogether. We also implement a 'forced choice' scorer.

Load

We probably actually want to start the pipeline from the betas rather than loading from pickle. to be continued...

new code

whole brain

Now let's scale that up to the full dataset.

masks

We get the same file this time, but we apply a mask before doing learning.

Next to do:

  1. Consider re-implementing LinearSVM and getting similarity rather than this probability measure
  2. Visualize for all the other values.

Now we've decoupled the learner from the framework where we apply it, we can pass in any arbitrary learner. It just has to output:

  1. predictions
  2. prediction confidence
  3. the classifier itself

Re-implementing v3

This is a useful sanity check, becuase it's helpufl to try and explain why we're doing so much better than rpeviously.

Formerly this was .85, .95, which was remarkable--we've now pretty clearly reproduced the result we were getting using nltools..

However, after I updated it based on documentation here (https://nilearn.github.io/modules/generated/nilearn.decoding.Decoder.html) it performed less well. Not sure why the classifier is performing less well in this setup than the former one - perhaps hte sort of normalization that I am doing on each side.

v3 mask

Apply the main analysis, looping through masks

Contrast between CorrectStop and CorrectGo...

Test 2

So actually, probably neural activity explains additional variance in a very straightforward way.

We found some relationships with FFQ. If these hold, it would be interesting to see support from them.

sooo...let's get a record of all the items from test 1 that we want to try running...

PES-related

Discriminability

whole-brain:

masked: